Anthropic Releases Claude 3.7 Sonnet With Extended Thinking Mode

Anthropic released Claude 3.7 Sonnet on February 24, 2025, introducing an extended thinking capability that allows the model to reason through complex problems before producing a response. The release marks the first Claude model to combine standard response generation with a visible, controllable chain-of-thought reasoning process in a single model.

What Happened

Anthropic published Claude 3.7 Sonnet to its Claude.ai platform and made it available via API to developers. The model can operate in two modes: a standard mode that generates responses at conventional speeds, and an extended thinking mode in which the model works through intermediate reasoning steps before delivering a final answer.

In extended thinking mode, users and developers can observe the model's reasoning process and set a token budget that controls how long the model deliberates. Anthropic said the feature is designed to improve performance on tasks requiring multi-step reasoning, including mathematics, coding, and scientific analysis.

Performance and Benchmarks

Anthropic released benchmark results alongside the model. Claude 3.7 Sonnet with extended thinking scored 96.2 percent on the MATH benchmark for mathematical reasoning and 70.3 percent on the GPQA Diamond benchmark, which tests graduate-level scientific knowledge. On the SWE-bench Verified coding benchmark, the model achieved 62.3 percent, which Anthropic described as the highest score reported for any publicly available model on that test at the time of release.

In standard mode, without extended thinking enabled, Claude 3.7 Sonnet also outperformed its predecessor, Claude 3.5 Sonnet, across several categories including instruction following and coding tasks, according to Anthropic's internal evaluations.

Pricing and Availability

Anthropic priced Claude 3.7 Sonnet at three dollars per million input tokens and fifteen dollars per million output tokens through the API, consistent with the pricing of Claude 3.5 Sonnet. Extended thinking tokens are billed at the same output token rate.

The model is available immediately on Claude.ai for Pro, Team, and Enterprise subscribers. API access is open to all registered developers. Anthropic also indicated that Claude 3.7 Sonnet is accessible through Amazon Web Services Bedrock and Google Cloud Vertex AI, where Anthropic has existing distribution agreements.

Background

Anthropic, founded in 2021 by former OpenAI researchers including CEO Dario Amodei and President Daniela Amodei, has positioned Claude as a safety-focused alternative to models from OpenAI, Google, and Meta. The company has raised more than seven billion dollars in funding, with major investments from Google and Amazon.

The release follows a period of rapid development in visible chain-of-thought reasoning across the industry. OpenAI introduced its o1 and o3 model series with similar reasoning capabilities in 2024 and early 2025. Google DeepMind released Gemini 2.5 Pro with enhanced reasoning features in April 2025. Claude 3.7 Sonnet represents Anthropic's direct entry into this category of models that expose intermediate reasoning steps to users.

What It Means in Practice

Developers using the API can set a maximum token budget for the thinking process, allowing them to balance response latency against reasoning depth depending on the application. Anthropic noted that extended thinking is particularly applicable to agentic workflows, in which an AI model must complete sequences of actions or decisions autonomously rather than answering a single question.

For enterprise customers, the ability to observe reasoning steps may assist with auditing model outputs and identifying errors in complex workflows, according to Anthropic's technical documentation accompanying the release.

Anthropic has indicated it plans to release additional models in the Claude 3.7 family, including a Claude 3.7 Opus variant, at a later date.